WG overlap: services and control

[Jeff Chase]

Since one of the goals for the upcoming meeting is to work on cross-WG
issues, I have posted some notes on the Services wiki about our tendrils
to the control WG.

http://groups.geni.net/geni/wiki/GeniServices/Tendrils/Fcf.html

This interaction in particular raises some basic issues about what
powers the "facility" actually has with respect to its components, and
how the services reflect those powers.  These issues generated a good
discussion at the meeting at BBN a couple of weeks ago, and we'll need
to make sure we are all on the same page before we go forward.   We are
interested in hearing perspectives from others in both working groups.

Jeff


_______________________________________________
control-wg mailing list
control-wg@...
http://lists.geni.net/mailman/listinfo/control-wg

[Schwab, Stephen]

Hi Jeff,
I'm just reading your points from the wiki on Tendrils now.
My two cents:

1. Slice Construction -- coallocation would be nice.  To date, other
testbeds (Emulab, DETER) have managed the task of swapping in large
experiments by manual human interactions -- e.g. during the inevitable
crunch before a deadline, coordination is done among the community of
users, aided by testbed operators with visibility into the current
allocation of nodes.  As someone who has jumped in to act as the
benevolent dictator to do scheduling, it works well in a small,
tight-knit community (the EMIST research project was an example), but is
non-scalable.  Coallocation would seem to be a nice feature -- but I'm
afraid of how hard this problem could really be.  In the most general
case, each experimenter would have a complex set of conditions to
satisfy, and the scheduler would need to jointly optimize assignment of
resources so that there was "fairness" in allocating the components to
the slices on behalf of the entire set of outstanding requests from
experimenters.  Maybe you are thinking of something simpler?

2. Slice fidelity.  I think this will need to be done as extra
side-measurements that are collected in parallel with the experiment, to
act as witnesses for the ground truth experienced by the components and
the slices running on those components during the experiment.  That is,
I don't think we can guarantee a certain level of fidelity or isolation
is perfect, and what it means to be good enough is relative to the
experiment in question.  We will need to collect enough
side-measurements, and provide the right tools and analysis to the
experimenter to let them calibrate their experiment against the ground
truth of whatever else was going on.  If we do things right, most
experiments will experience perturbations from other experiments, or the
background load on shared networks, that is minor noise.  Our goal
should be to enable researchers to calculate bounded error bars, and
know how to control the (minor) errors introduced by using GENI.
Consider also that large scale systems and networks have real failures
-- so GENI experiments will need to tolerate failure of underlying
components, or they won't scale.

3. Policy control -- Sorry, this just makes my head hurt. :-)

--Steve

-----Original Message-----
From: control-wg-bounces@... [mailto:control-wg-bounces@...]
On Behalf Of Jeff Chase
Sent: Friday, February 29, 2008 4:08 PM
To: control-wg@...; services
Subject: [cwg] WG overlap: services and control

Since one of the goals for the upcoming meeting is to work on cross-WG
issues, I have posted some notes on the Services wiki about our tendrils

to the control WG.

http://groups.geni.net/geni/wiki/GeniServices/Tendrils/Fcf.html

This interaction in particular raises some basic issues about what
powers the "facility" actually has with respect to its components, and
how the services reflect those powers.  These issues generated a good
discussion at the meeting at BBN a couple of weeks ago, and we'll need
to make sure we are all on the same page before we go forward.   We are
interested in hearing perspectives from others in both working groups.

Jeff


_______________________________________________
control-wg mailing list
control-wg@...
http://lists.geni.net/mailman/listinfo/control-wg

_______________________________________________
control-wg mailing list
control-wg@...
http://lists.geni.net/mailman/listinfo/control-wg